CyberMCP — a free, open-source pentesting & bug bounty MCP server for AI agents

CyberMCP turns any AI agent (Claude, OpenCode, Cursor, Cline, RooCode, Continue, etc.) into a capable authorized pentest/bug-bounty assistant. It enforces engagement scope, keeps a full audit trail, and provides safe, detection-first recon + vulnerability checks, plus reporting.

Features

• Scope-first safety — scope.yaml per engagement (domains/CIDRs, wildcards, out-of-scope, rate/concurrency). Every network action is guard-checked.

• Full audit trail — append-only audit.jsonl for every action, tool run, finding and scope bypass (if any).

• Recon — DNS lookup, subdomain enum (crt.sh + DNS wordlist), async TCP port scan, HTTP probing, WAF hints.

• Web analysis — flexible HTTP client, content discovery, parameter discovery, tech fingerprinting, security headers, TLS inspection, CORS.

• Vuln detection (safe) — exposed .git/.env/backups, open redirect, SQLi error-based, reflected XSS, LFI, sensitive paths, auth gaps.

• Payload & wordlist generation — context-aware payloads tailored to detected tech/DB; mines live pages to build custom wordlists; merge/dedup.

• Orchestration pipelines — full_recon_pipeline, web_assessment_pipeline, vuln_scan_pipeline stage results and can auto-record findings.

• Crypto & secrets — JWT decode + weak-secret detection, hash identification, encoding/decoding toolbox.

• External CLI wrappers — nmap/nuclei/ffuf/sqlmap/subfinder/httpx (auto-detected, scope-checked, time-limited, pure-Python fallbacks exist).

• Findings & reporting — add/update/list/findings; export JSON; generate professional Markdown + HTML pentest reports.

• Embedded knowledge — OWASP Top 10 2021, methodology, web cheatsheet, reporting standards, tool cheatsheets, scoping/ethics.

• MCP prompts — reusable workflows (full_recon, web_assessment, vulnerability_validation, report_writer).

Related MCP server: BugBounty MCP Server

Quick Start

1) Install

From PyPI (once published):

uvx cyber-mcp

From source (works today):

git clone https://github.com/Dacraezy1/CyberMCP
cd CyberMCP
pip install .
cyber-mcp

2) Configure with an MCP client (examples)

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "cyber-mcp": {
      "command": "uvx",
      "args": ["cyber-mcp"]
    }
  }
}

Cursor (mcp.json)

{
  "mcpServers": {
    "cyber-mcp": {
      "command": "uvx",
      "args": ["cyber-mcp"]
    }
  }
}

OpenCode / Cline / RooCode

Use command: uvx, args: ["cyber-mcp"].

3) Use in an agent (typical flow)

1. Engage: init_engagement(name="demo", in_scope_domains=["example.com"], rules="Authorized test only.")

2. Recon: full_recon_pipeline(target="example.com")

3. Assess: web_assessment_pipeline(url="https://example.com")

4. Review: list_findings()

5. Report: generate_report(fmt="both")

All actions are scope-enforced and audited.

Development

Clone & run in dev mode

git clone https://github.com/Dacraezy1/CyberMCP
cd CyberMCP
uv venv
source .venv/bin/activate
uv pip install -e ".[dev]"
cyber-mcp

Tests

pytest -v

Safety Notes

• Detection-first: built-in checks are non-destructive (error-based signatures, reflection canaries, metadata reads). They don't exploit, modify data, or exfiltrate bulk secrets.

• Scope is king: if a target isn't in your engagement scope, tools refuse it. Private/reserved IPs are blocked by default.

• External tools: wrappers add timeouts, scope checks and restricted flags. Increase risk only with written authorization.

• Auditability: every action goes to engagements/<name>/audit.jsonl. Keep it intact for post-test review.

• No warranty: provided as-is, for authorized security testing only. Users are responsible for compliance with applicable laws and program rules.

License

MIT License

Contributing

Contributions welcome. Please:

• Keep changes consistent with existing style and architecture (scope guard + audit first).

• Add/extend tests for new tools.

• Document new tools/prompts in README and knowledge.

• Maintain safety: scope checks on all network tools.